EstevaoNaval/visualheist-base-onnx

VisualHeist base ONNX (fp16)

Production ONNX Runtime bundle for VisualHeist-base object detection (<OD>): three graphs plus tokenizer and preprocessor config.

Files

File	Role
encoder_fp16.onnx + .data	Vision + language encoder (once per page)
decoder_init_fp16.onnx + .data	Decoder first token (beam batch)
decoder_step_fp16.onnx + .data	Decoder steps with KV cache
tokenizer.json	Bart tokenizer for decoding generated ids
preprocessor_config.json	Image size (768×768), mean, std

Runtime

• ONNX Runtime on CPU (CPUExecutionProvider)
• Beam search in Python (chemtables.visualheist), not inside ONNX
• Cached by chemtables at ~/.cache/pdf2chemicals/models/visualheist-onnx on first PDF crop (override with VISUALHEIST_ONNX_DIR or PDF2CHEMICALS_DATA_DIR)

Export

Graphs are produced from local safetensors via scripts/export_visualheist_onnx.py in the pdf2chemicals repository.